Claude’s 1M context goes GA at standard pricing

If you've ever watched an AI agent try to debug a complex system or review a massive pull request, you know the frustration: just as things get interesting, context compaction kicks in, earlier details vanish, and the agent starts hallucinating in circles.

Anthropic just made the 1M token context window generally available for Claude Opus 4.6 and Sonnet 4.6. The real kicker isn't just the capacity—it's that they completely killed the long-context pricing premium.

Standard pricing all the way to 1M

Until now, running frontier models at deep context often felt like paying luxury tax. Anthropic has leveled the billing structure across the board.

Opus 4.6 stays at $5 per million input tokens and $25 per million output tokens. Sonnet 4.6 stays at $3/$15. A 900K-token prompt is billed at the exact same token rate as a 9K query, with no context multiplier and full standard rate limits intact.

Media payloads also received a substantial bump. You can now pass up to 600 images or PDF pages per request, a sixfold increase from the prior 100-page ceiling.

For API users, the beta header requirement is gone—requests exceeding 200K tokens expand automatically without breaking existing code. The update is live natively on Claude Platform and rolling through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Meanwhile, Claude Code users on Max, Team, and Enterprise plans will see Opus 4.6 default to 1M context out of the box.

Solving agent amnesia

A massive context window is useless if the model hallucinates what it read 400 pages ago. Anthropic claims Opus 4.6 scored 78.3% on the MRCR v2 benchmark, which they report as the top score among frontier models at that length.

Early feedback from builders highlights immediate workflow changes. The team behind Devin noted that large code diffs previously overflowed the 200K limit, forcing chunked context that missed cross-file dependencies. Feeding full diffs into 1M context produced cleaner reviews with fewer tokens spent on coordination.

Other developers reported that lifting context constraints actually reduced their overall token spend. When an agent doesn't need to summarize, re-read, and verify intermediate steps, it stops running in circles.

Why it matters

We think the pricing move is a deliberate squeeze on the rest of the frontier market. By eliminating the long-context surcharge, Anthropic is turning 1M tokens from a premium, use-only-when-desperate feature into standard production infrastructure.

For engineers, this dramatically simplifies the software harness around autonomous agents. Complex RAG pipelines, lossy vector chunking, and aggressive memory-clearing wrappers can finally take a back seat. If you're building legal-document workflows, system incident responders, or complex codebase assistants, trying 1M context is an immediate no-brainer.

That said, flat pricing doesn't mean free compute. When a single multi-turn agent loop casually ingests 500K tokens per step, API bills can escalate rapidly without strict usage guardrails. The window is wider, but the cost discipline still falls squarely on you.